Image Forgery Localization Based On Deep Learning Algorithm

With the widespread use of image editing software,non-expert users can also easily edit digital images,which poses a certain threat to the integrity and authenticity of images.It is undeniable that the progress of image technology has brought convenience to people’s lives,but images are maliciously tampered with and used to confuse the public and reverse black and white,which will produce strong destructive power,affecting personal reputation,public opinion orientation and even national security.Therefore,the research of image tamper detection technology is of great practical significance.This thesis focuses on image stitching tampering detection and tampering area positioning,and the main work content is as follows:Firstly,a dual-stream image forgery detection algorithm combined with noise inconsistency is proposed,and the noise feature inconsistency of the tampered region image and the non-tampered region image are used to locate the tampered region.In addition to feature extraction for the semantic information of the tampered image,the dual-stream network combines the noise branch to extract the feature of the noise information of the tampered image.Small tampering targets have always been a troublesome problem,and small tampering targets will lead to fewer inconsistent features in the entire tampered image,so it is easy to be ignored and difficult to detect in the process of feature extraction.In view of this situation,while looking for semantic information inconsistencies through semantic branching,the algorithm in this thesis synchronously learns the inconsistency of noise features in combination with noise flow input,combines the inconsistency of image noise details with the inconsistency of semantic features such as image color,brightness,contrast,edge,etc.,and learns as many different features of the tampering region and non-tampering region as possible,and obtains the complete tampering target.Compared with the detection effect of the single-stream model in CASIA dataset,the dual-stream image stitching tampering positioning algorithm has been significantly improved.Secondly,an image tamper detection algorithm based on multi-scale integrated attention mechanism is proposed.Based on the previous article,the algorithm further explores how to better detect the tampering area of small targets.The combination of RGB stream and noise flow brings more parameters to the previous algorithm,which increases the difficulty of training,and the network focuses on the noise inconsistent tamper image,which does not bring good results for other categories,such as JPEG compression type tamper image.The attention mechanism can improve the attention to the tampered area by weighting the feature map as a whole,and retain the detailed information of the tampered area in this way,improve the detection effect,and be more versatile,and the attention mechanism is relatively lightweight in a single-stream network.The algorithm introduces two attention networks,on the one hand,the location attention is weighted for the location area of the feature map,and on the other hand,the local information of the feature map can be carefully learned through multi-scale division to ensure that the tampering area is not missed.The channel attention is deep in the network,weighted according to the semantic characteristics of the image content,and the semantic information of the tampering target at the edge is obtained on the basis of non-leakage.Through the weighting of these two sets of attention mechanisms in the process of feature map processing,the algorithm achieves better detection effect.Finally,a large number of ablation experiments and comparative experiments are carried out in this thesis.Ablation experiments are conducted for the improvement and adjustment of the algorithm,comparative experiments are conducted for the advancement of the algorithm,and finally robustness analysis experiments are conducted to evaluate the tampering detection effect from qualitative and quantitative perspectives,verifying that the improved algorithm can obtain better detection results.